基于近红外光谱技术的大豆种子老化级别快速鉴别方法研究

大豆种子容易发生老化并丧失活力,大豆种子活力检测对目前农业生产具有重要意义。以2020年收获的大豆种子为样本进行人工老化试验,老化时间设置为1、2、3、4、5、6d,以未老化的种子作为对照组,每个老化等级30个样本。扫描获取全部210条近红外光谱数据,以4：1的比例划分样本集。对原始光谱数据建立BP网络模型1,再分别采取多元散射校正和标准正太变量变换对原始光谱进行预处理,建立模型2,模型3。比较3种模型可以发现预处理技术能缩短模型迭代时间,同时可以消除部分噪声,提高模型预测能力,且经过标准正太变量变换处理后的模型结果较优,由于预处理后的数据维度并未发生变化,模型的迭代时间较长,不利于实际应用。因此采取主成分分析、连续投影法、竞争自适应重加权法对经过标准正太变换后的数据进行特征波长变量提取,将光谱数据由原来的1845维降到10维、23维和150维。对经过特征波长变量提取后的数据分别建立BP网络模型,得到模型4、5、6。综合分析上述六种模型,最终建立了150输入、10个隐层、7个输出的神经网络鉴别模型6,其分类准确率达到93.43%,迭代时间2.25s可以较好实现对七类不同老化级别的大豆种子快速、无损鉴别。

Rapid Identification of Soybean Seed Aging Grade Based on Near Infrared Spectroscopy

Soybean seeds are easy to age and lose their vigor. The detection of soybean seed vigor is of great significance to current agricultural production. Artificial aging test of soybean seeds harvested in 2020,and the aging time was set as 1,2,3,4,5, and 6 days. Non-aging seeds were used as the control group, with 30 samples for each aging grade. All 210 NIR spectral data were obtained by scanning, and the sample set was divided by 4:1 ratio. 168 samples were obtained as the modeling set, and 42 samples were obtained as the prediction set. BP network model 1 was established for the original spectral data, and then multiple scattering correction and standard positive ternary transformation were used to preprocess the original spectrum respectively, and model 2 and model 3 were established. Comparison model 1, 2, 3 can be found on the original spectral preprocessing, model iteration can shorten the time and can eliminate some noise at the same time, improve the model prediction ability, and Standard Normal Variate make variable transformation is superior to the multiple scattering correction. Since the data dimensions after preprocessing have not changed, the iteration time of the model is long, which is not conducive to practical application. Therefore, principal component analysis, continuous projection method and competitive adaptive reweighting method were adopted to extract characteristic wavelength variables from the data after standard orthographic transformation, and the spectral data were reduced from the original 1845 dimension to 10, 23 and 150 dimensions. BP network models are established for the data extracted from the characteristic wavelength variables, and models 4, 5, and 6 are obtained. By comprehensive analysis of the above six models, a neural network identification model with 150 inputs, 10 hidden layers and 7 outputs was established. The classification accuracy reached 93.43%, and the iteration time was 2.25s, which could effectively realize the rapid and nondestructive identification of seven types of soybean seeds with different aging grades.